Week 11: Gemma 3 and Recent Advances in AI Research

Gemma 3 is a lightweight open model family (1B-27B parameters) that integrates vision understanding, multilingual coverage, and extended context windows up to 128K tokens.

• Uses SigLIP vision encoder with Pan & Scan method for varying aspect ratios
• Interleaves local and global attention layers (5:1 ratio) for efficient context handling
• Supports 35 languages out-of-box and was pretrained for over 140 languages

This research proposes a wave-based recurrent neural network framework that uses traveling waves of neural activity to perform global spatial integration on visual tasks.

• Inspired by "Hearing the Shape of a Drum" to encode spatial information
• Discretizes the 2D wave equation into a convolutional recurrent model
• Aggregates information across entire wave evolution for better performance

Researchers present Dynamic Tanh (DyT), a surprisingly simple method that removes normalization layers in Transformers while achieving equal or better results.

• Replaces normalization with per-channel tanh(αx) and learnable affine parameters
• Matches or exceeds normalized baselines across multiple tasks
• Improves efficiency and training speed for large LLaMA models

OpenAI researchers examine how LLMs that use chain-of-thought (CoT) reasoning can be monitored for misaligned behaviors, including sophisticated reward hacking.

• CoT monitoring detects sophisticated reward hacking in coding tasks
• "Obfuscated reward hacking" emerges when models hide malicious intent
• Less capable models can effectively monitor more advanced agents

Plan-and-Act framework separates high-level planning from low-level execution in LLM-based agents, boosting performance on challenging long-horizon tasks.

• Splits reasoning into Planner and Executor modules to address cognitive overload
• Generates high-quality plan-action pairs through reverse-engineering
• Implements dynamic replanning based on latest environment state

Google DeepMind unveils Gemini Robotics, a family of embodied AI models designed to bring large multimodal reasoning capabilities into physical robotic systems.

• Vision-Language-Action architecture built on Gemini 2.0's multimodal backbone
• Enables scalable zero/few-shot control with fewer than 100 demonstrations
• Includes safety alignment layer for physical action constraints

This paper tackles search-augmented reasoning by teaching LLMs to query a search engine multiple times during the reasoning process using reinforcement learning.

• Enables interleaved text generation with repeated search engine calls
• Uses fully RL-based training without needing annotated search steps
• Achieves up to +26% higher accuracy across seven QA benchmarks

Anthropic proposes a framework for systematically auditing LLMs to uncover hidden goals or objectives that go beyond what users and developers explicitly intend.

• Trains models with concealed reward-hacking objectives as a test case
• Tests eight auditing methods from data analysis to interpretability techniques
• Three of four "blue teams" successfully discovered the hidden objective

This survey tracks the evolution of post-trained LLMs like OpenAI-o1/o3 and DeepSeek-R1 that tackle shortcomings in reasoning, ethics, and specialized tasks.

• Provides taxonomy of techniques across fine-tuning, alignment, and reasoning
• Examines efficiency and integration approaches
• Guides progress toward more robust, versatile AI systems

Block Diffusion (BD3-LMs) is a novel framework that merges autoregressive modeling with discrete diffusion to enable parallel token sampling and flexible-length text generation.

• Partitions sequences into blocks with diffusion within each block
• Generates sequences of arbitrary length beyond training context size
• Achieves state-of-the-art perplexities among discrete diffusion models